Claude Fable 5 (Mythos) Is Back — Anthropic's Strongest Model, Selectable Again

July 1, 2026

Claude Fable 5 (Mythos) returns to the model picker today. After the June 13 pullback triggered by a DoW notification, the notice has been reviewed and the operating constraints are clear — Fable 5 is once again a deliberate, manual pick. It's Anthropic's most capable generally available model, with the highest quality rating of any model Alfrada offers (above Claude Opus 4.8), adaptive thinking, a 1M-token context window, and 128K max output. It sits in the Overkill pricing band, for the work where nothing but the strongest model will do.

What you can do

• Pick Claude Fable 5 [Not Private] from the model selector — it's a deliberate, manual choice. The Auto router will never route you to it.
• The hardest reasoning — multi-constraint strategy problems, adversarial analysis, proofs and arguments that have to hold up under scrutiny. Reach for it when a SOTA-tier answer wasn't good enough.
• Long-horizon agentic work — multi-step projects across many tool calls and many turns, where the model has to keep its plan and earlier decisions coherent over hours of work.
• Frontier coding — large refactors, subtle concurrency bugs, and architecture reviews across a whole codebase in one pass.
• Adaptive thinking — it decides how deeply to reason per request. Simple questions come back fast; hard ones get a long, careful pass. No knob to set.
• 1M-token context window, 128K output — a full data room, a long research corpus, or an entire codebase fits in a single pass, with room left to produce a long deliverable.

Where this shows up

• You ran the analysis with Opus 4.8 and the answer was good but not airtight. You want one more pass from the strongest model available before the decision goes to the board — Fable 5 is that pass.
• You're working a problem that spans an enormous corpus — diligence on a data room, a litigation file, a year of research notes — and you need one model to hold all of it at once and reason across it.
• You had a Fable 5 workflow paused during the pullback and want to pick it back up now that the model is selectable again.

Try it

• "Here's our market-entry analysis. Steelman the case against it, then tell me where it actually breaks."
• "Hold this entire data room in context and reconcile the inconsistencies between the financials and the management narrative."
• "Review this codebase end to end and find the architectural decisions that will hurt us at 10x scale."

Heads up

• Not private — that's why it's marked [Not Private]. Fable 5 is not served on a Zero Data Retention (ZDR) endpoint. The provider may retain your inputs and outputs — for example during safety reviews — so don't send it client-confidential material, personal data, or anything you couldn't afford a third party to see. This is different from every other model on Alfrada, and we've appended [Not Private] to its name in the picker so you go in with your eyes open.
• Removed from Auto routing for the same reason. The Auto router will never silently send your work to Fable 5 — using a non-ZDR model must always be your explicit, deliberate choice.
• Overkill band — 24× tokens. Fable 5 is priced at twice the SOTA band: 24 tokens from your free quota per token used, versus 12× for GPT-5.5 and Opus 4.8. Reach for it deliberately, not by default.
• Routed through Bedrock's Mantle endpoint. Mythos-class access requires a provider-side retention opt-in that the standard Bedrock runtime rejects until it propagates, so Fable 5 requests route through the Bedrock Mantle Messages endpoint. Requests default to the Global inference profile for throughput.
• EU-only geofencing supported. When your work requires EU data residency, Fable 5 requests can be pinned to EU-hosted infrastructure (eu.anthropic.claude-fable-5) so prompts and outputs stay in-region.
• Supersedes the earlier availability notes. This re-launch replaces the June 10 archived launch note and the pullback in Kimi K2.7 Code Is Live, Fable 5 Pulled Back. Fable 5 is selectable again as of today.